Results for
Since R2024b, a Levenberg–Marquardt solver (TrainingOptionsLM) was introduced. The built‑in function trainnet now accepts training options via the trainingOptions function (https://www.mathworks.com/help/deeplearning/ref/trainingoptions.html#bu59f0q-2) and supports the LM algorithm. I have been curious how to use it in deep learning, and the official documentation has not provided a concrete usage example so far. Below I give a simple example to illustrate how to use this LM algorithm to optimize a small number of learnable parameters.
For example, consider the nonlinear function:
y_hat = @(a,t) a(1)*(t/100) + a(2)*(t/100).^2 + a(3)*(t/100).^3 + a(4)*(t/100).^4;
It represents a curve. Given 100 matching points (t → y_hat), we want to use least squares to estimate the four parameters a1–a4.
t = (1:100)';
y_hat = @(a,t)a(1)*(t/100) + a(2)*(t/100).^2 + a(3)*(t/100).^3 + a(4)*(t/100).^4;
x_true = [ 20 ; 10 ; 1 ; 50 ];
y_true = y_hat(x_true,t);
plot(t,y_true,'o-')
- Using the traditional lsqcurvefit-wrapped "Levenberg–Marquardt" algorithm:
x_guess = [ 5 ; 2 ; 0.2 ; -10 ];
options = optimoptions("lsqcurvefit",Algorithm="levenberg-marquardt",MaxFunctionEvaluations=800);
[x,resnorm,residual,exitflag] = lsqcurvefit(y_hat,x_guess,t,y_true,-50*ones(4,1),60*ones(4,1),options);
x,resnorm,exitflag
- Using the deep-learning-wrapped "Levenberg–Marquardt" algorithm:
options = trainingOptions("lm", ...
InitialDampingFactor=0.002, ...
MaxDampingFactor=1e9, ...
DampingIncreaseFactor=12, ...
DampingDecreaseFactor=0.2,...
GradientTolerance=1e-6, ...
StepTolerance=1e-6,...
Plots="training-progress");
numFeatures = 1;
layers = [featureInputLayer(numFeatures,'Name','input')
fitCurveLayer(Name='fitCurve')];
net = dlnetwork(layers);
XData = dlarray(t);
YData = dlarray(y_true);
netTrained = trainnet(XData,YData,net,"mse",options);
netTrained.Layers(2)
classdef fitCurveLayer < nnet.layer.Layer ...
& nnet.layer.Acceleratable
% Example custom SReLU layer.
properties (Learnable)
% Layer learnable parameters
a1
a2
a3
a4
end
methods
function layer = fitCurveLayer(args)
arguments
args.Name = "lm_fit";
end
% Set layer name.
layer.Name = args.Name;
% Set layer description.
layer.Description = "fit curve layer";
end
function layer = initialize(layer,~)
% layer = initialize(layer,layout) initializes the layer
% learnable parameters using the specified input layout.
if isempty(layer.a1)
layer.a1 = rand();
end
if isempty(layer.a2)
layer.a2 = rand();
end
if isempty(layer.a3)
layer.a3 = rand();
end
if isempty(layer.a4)
layer.a4 = rand();
end
end
function Y = predict(layer, X)
% Y = predict(layer, X) forwards the input data X through the
% layer and outputs the result Y.
% Y = layer.a1.*exp(-X./layer.a2) + layer.a3.*X.*exp(-X./layer.a4);
Y = layer.a1*(X/100) + layer.a2*(X/100).^2 + layer.a3*(X/100).^3 + layer.a4*(X/100).^4;
end
end
end
The network is very simple — only the fitCurveLayer defines the learnable parameters a1–a4. I observed that the output values are very close to those from lsqcurvefit.
I’d like to take a moment to highlight the great contributions of one of our community members, @Paul, who is fast approaching an impressive 5,000 reputation points!
Paul has built his reputation the best way possible - by generously sharing his knowledge and helping others. Over the last few years, he’s provided thoughtful and practical answers to hundreds of questions, making life a little easier for learners and experts alike.
Reputation points are more than just numbers here - they represent the trust and appreciation of the community. Paul’s upcoming milestone is a testament to his consistency, expertise, and willingness to support others.
Please join me in recognizing Paul's contributions and impact on the MATLAB Central community.
Function Syntax Design Conundrum
As a MATLAB enthusiast, I particularly enjoy Steve Eddins' blog and the cool things he explores. MATLAB's new argument blocks are great, but there's one frustrating limitation that Steve outlined beautifully in his blog post "Function Syntax Design Conundrum": cases where an argument should accept both enumerated values AND other data types.
Steve points out this could be done using the input parser, but I prefer having tab completions and I'm not a fan of maintaining function signature JSON files for all my functions.
Personal Context on Enumerations
To be clear: I honestly don't like enumerations in any way, shape, or form. One reason is how awkward they are. I've long suspected they're simply predefined constructor calls with a set argument, and I think that's all but confirmed here. This explains why I've had to fight the enumeration system when trying to take arguments of many types and normalize them to enumerated members, or have numeric values displayed as enumerated members without being recast to the superclass every operation.
The Discovery
While playing around extensively with metadata for another project, I realized (and I'm entirely unsure why it took so long) that the properties of a metaclass object are just, in many cases, the attributes of the classdef. In this realization, I found a solution to Steve's and my problem.
To be clear: I'm not in love with this solution. I would much prefer a better approach for allowing variable sets of membership validation for arguments. But as it stands, we don't have that, so here's an interesting, if incredibly hacky, solution.
If you call struct() on a metaclass object to view its hidden properties, you'll notice that in addition to the public "Enumeration" property, there's a hidden "Enumerable" property. They're both logicals, which implies they're likely functionally distinct. I was curious about that distinction and hoped to find some functionality by intentionally manipulating these values - and I did, solving the exact problem Steve mentions.
The Problem Statement
We have a function with an argument that should allow "dual" input types: enumerated values (Steve's example uses days of the week, mine uses the "all" option available in various dimension-operating functions) AND integers. We want tab completion for the enumerated values while still accepting the numeric inputs.
A Solution for Tab-Completion Supported Arguments
Rather than spoil Steve's blog post, let me use my own example: implementing a none() function. The definition is simple enough tf = ~any(A, dim); but when we wrap this in another function, we lose the tab-completion that any() provides for the dim argument (which gives you "all"). There's no great way to implement this as a function author currently - at least, that's well documented.
So here's my solution:
%% Example Function Implementation
% This is a simple implementation of the DimensionArgument class for implementing dual type inputs that allow enumerated tab-completion.
function tf = none(A, dim)
arguments(Input)
A logical;
dim DimensionArgument = DimensionArgument(A, true);
end
% Simple example (notice the use of uplus to unwrap the hidden property)
tf = ~any(A, +dim);
end
I like this approach because the additional work required to implement it, once the enumeration class is initialized, is minimal. Here are examples of function calls, note that the behavior parallels that of the MATLAB native-style tab-completion:
%% Test Data
% Simple logical array for testing
A = randi([0, 1], [3, 5], "logical");
%% Example function calls
tf = none(A, "all"); % This is the tab-completion it's 1:1 with MATLABs behavior
tf = none(A, [1, 2]); % We can still use valid arguments (validated in the constructor)
tf = none(A); % Showcase of the constructors use as a default argument generator
How It Works
What makes this work is the previously mentioned Enumeration attribute. By setting Enumeration = false while still declaring an enumeration block in the classdef file, we get the suggested members as auto-complete suggestions. As I hinted at, the value of enumerations (if you don't subclass a builtin and define values with the someMember (1) syntax) are simply arguments to constructor calls.
We also get full control over the storage and handling of the class, which means we lose the implicit storage that enumerations normally provide and are responsible for doing so ourselves - but I much prefer this. We can implement internal validation logic to ensure values that aren't in the enumerated set still comply with our constraints, and store the input (whether the enumerated member or alternative type) in an internal property.
As seen in the example class below, this maintains a convenient interface for both the function caller and author the only particuarly verbose portion is the conversion methods... Which if your willing to double down on the uplus unwrapping context can be avoided. What I have personally done is overload the uplus function to return the input (or perform the identity property) this allowss for the uplus to be used universally to unwrap inputs and for those that cant, and dont have a uplus definition, the value itself is just returned:
classdef(Enumeration = false) DimensionArgument % < matlab.mixin.internal.MatrixDisplay
%DimensionArgument Enumeration class to provide auto-complete on functions needing the dimension type seen in all()
% Enumerations are just macros to make constructor calls with a known set of arguments. Declaring the 'all'
% enumeration member means this class can be set as the type for an input and the auto-completion for the given
% argument will show the enumeration members, allowing tab-completion. Declaring the Enumeration attribute of
% the class as false gives us control over the constructor and internal implementation. As such we can use it
% to validate the numeric inputs, in the event the 'all' option was not used, and return an object that will
% then work in place of valid dimension argument options.
%% Enumeration members
% These are the auto-complete options you'd like to make available for the function signature for a given
% argument.
enumeration(Description="Enumerated value for the dimension argument.")
all
end
%% Properties
% The internal property allows the constructor's input to be stored; this ensures that the value is store and
% that the output of the constructor has the class type so that the validation passes.
% (Constructors must return the an object of the class they're a constructor for)
properties(Hidden, Description="Storage of the constructor input for later use.")
Data = [];
end
%% Constructor method
% By the magic of declaring (Enumeration = false) in our class def arguments we get full control over the
% constructor implementation.
%
% The second argument in this specific instance is to enable the argument's default value to be set in the
% arguments block itself as opposed to doing so in the function body... I like this better but if you didn't
% you could just as easily keep the constructor simple.
methods
function obj = DimensionArgument(A, Adim)
%DimensionArgument Initialize the dimension argument.
arguments
% This will be the enumeration member name from auto-completed entries, or the raw user input if not
% used.
A = [];
% A flag that indicates to create the value using different logic, in this case the first non-singleton
% dimension, because this matches the behavior of functions like, all(), sum() prod(), etc.
Adim (1, 1) logical = false;
end
if(Adim)
% Allows default initialization from an input to match the aforemention function's behavior
obj.Data = firstNonscalarDim(A);
else
% As a convenience for this style of implementation we can validate the input to ensure that since we're
% suppose to be an enumeration, the input is valid
DimensionArgument.mustBeValidMember(A);
% Store the input in a hidden property since declaring ~Enumeration means we are responsible for storing
% it.
obj.Data = A;
end
end
end
%% Conversion methods
% Applies conversion to the data property so that implicit casting of functions works. Unfortunately most of
% the MathWorks defined functions use a different system than that employed by the arguments block, which
% defers to the class defined converter methods... Which is why uplus (+obj) has been defined to unwrap the
% data for ease of use.
methods
function obj = uplus(obj)
obj = obj.Data;
end
function str = char(obj)
str = char(obj.Data);
end
function str = cellstr(obj)
str = cellstr(obj.Data);
end
function str = string(obj)
str = string(obj.Data);
end
function A = double(obj)
A = double(obj.Data);
end
function A = int8(obj)
A = int8(obj.Data);
end
function A = int16(obj)
A = int16(obj.Data);
end
function A = int32(obj)
A = int32(obj.Data);
end
function A = int64(obj)
A = int64(obj.Data);
end
end
%% Validation methods
% These utility methods are for input validation
methods(Static, Access = private)
function tf = isValidMember(obj)
%isValidMember Checks that the input is a valid dimension argument.
tf = (istext(obj) && all(obj == "all", "all")) || (isnumeric(obj) && all(isint(obj) & obj > 0, "all"));
end
function mustBeValidMember(obj)
%mustBeValidMember Validates that the input is a valid dimension argument for the dim/dimVec arguments.
if(~DimensionArgument.isValidMember(obj))
exception("JB:DimensionArgument:InvalidInput", "Input must be an integer value or the term 'all'.")
end
end
end
%% Convenient data display passthrough
methods
function disp(obj, name)
arguments
obj DimensionArgument
name string {mustBeScalarOrEmpty} = [];
end
% Dispatch internal data's display implementation
display(obj.Data, char(name));
end
end
end
In the event you'd actually play with theres here are the function definitions for some of the utility functions I used in them, including my exception would be a pain so i wont, these cases wont use it any...
% Far from my definition isint() but is consistent with mustBeInteger() for real numbers but will suffice for the example
function tf = isint(A)
arguments
A {mustBeNumeric(A)};
end
tf = floor(A) == A
end
% Sort of the same but its fine
function dim = firstNonscalarDim(A)
arguments
A
end
dim = [find(size(A) > 1, 1), 0];
dim(1) = dim(1);
end
Hello MATLAB Central, this is my first article.
My name is Yann. And I love MATLAB.
I also love HTTP (i know, weird fetish)
So i started a conversation with ChatGPT about it:
gitclone('https://github.com/yanndebray/HTTP-with-MATLAB');
cd('HTTP-with-MATLAB')
http_with_MATLAB
I'm not sure that this platform is intended to clone repos from github, but i figured I'd paste this shortcut in case you want to try out my live script http_with_MATLAB.m
A lot of what i program lately relies on external web services (either for fetching data, or calling LLMs).
So I wrote a small tutorial of the 7 or so things I feel like I need to remember when making HTTP requests in MATLAB.
Let me know what you think
a = [ 1 2 3 ]
a =
1 2 3
b = [ 3; 5; 8]
b =
3
5
8
x = a + b
x =
4 5 6
6 7 8
9 10 11
Hi,
I have some problem, I want to upload my data that sample rate at 500HZ, every sevral seconds.
My data include 12 bytes, and it measure 500HZ, for example for 15 seconds I coolect 15*500*12 = 84KB.
Can I upload this data to ThingSpeak? It is possible to use with Free acount (I am student and this is my project)
How can help me..
I can not understand why Plot Browser was taken away in latest Matlab... I use Plot Browser all of the time! Having to find and click the particular line I want in a plot with a lot of lines is way less convenient than just selecting it in the Plot Browser. Also, being able to quickly hide/show multiple lines at once with the plot browser was so helpful in a lot of cases. Please bring Plot Browser back!!!! Please reply with support for this if you feel the same as I do!
In the latest Graphics and App Building blog article, documentation writer Jasmine Poppick modernized a figure-based bridge analysis app by replacing uicontrol with new UI components and uifigure, resulting in cleaner code, better layouts, and expanded functionality in R2025a.
https://blogs.mathworks.com/graphics-and-apps/2025/08/19/__from-uicontrol-to-ui-components
This article covers the following topics:
Why and when to move from uicontrol and figure to modern UI components and uifigure.
How to replace uicontrol objects with equivalent UI component functions (uicheckbox, uidropdown, uispinner, etc.).
How to update callback code to match new component properties and behaviors.
How to adopt new UI component types (like spinners) to simplify validation and improve usability.
How to configure existing components with modern options (sortable tables, auto-fitting columns, editable data).
How to apply visual styling with uistyle and addStyle to make apps more user-friendly.
How to use uigridlayout to create flexible, adaptive layouts instead of manually managing positions.
The benefits of switching from figure to uifigure for app-building workflows.
A full before-and-after example of modernizing an existing app with incremental, practical updates.
In our large open-source MATLAB Central community, there are many long-term excellent user groups. I really want to know why you have been using MATLAB for a long time, and what are its absolute advantages?
I have been using MATLAB for a long time, and there are several reasons for that:
- Fast ramp-up in unfamiliar domains: When I explore an unfamiliar application area or a new topic, MATLAB helps me quickly locate the canonical methods and example workflows. Its comprehensive, professional documentation — along with the related-topic links typically provided at the end of each page — makes it easy to get started intuitively and saves a lot of time that would otherwise be spent hunting for foundational knowledge across the web.
- A relatively cutting-edge yet reliable technical path: MATLAB’s many toolboxes evolve with the field. While updates aren’t always absolutely bleeding-edge, they generally offer approaches that balance modernity and proven reliability. This reduces the risk of wasting time on obscure or unstable algorithms and helps me follow a pragmatic, well-tested technical direction.
- Strong community and technical support: When I encounter a problem I first post on forums like MATLAB Answers and thoroughly investigate the issue myself. If I find a solution, I publish it to contribute back — which deepens my own understanding and helps others. If I can’t solve it alone, experienced community members often respond within hours. As a last resort, MathWorks’ official support is available and typically conducts an in-depth investigation into specific cases to help resolve the issue.
- ......
Also, most individuals have limited time and technical bandwidth, diving deeply into a single, narrow area can be hard to pull back from unless you are committed to that specific direction. For cutting‑edge, highly specialized research it’s often necessary to combine MATLAB with other languages (e.g., Python, C/C++) to go further.
There is a communication regarding "How can I set the text font style of a Data Cursor object interactively on a plot?". But I am not clear on the answer found in this link:
https://www.mathworks.com/matlabcentral/answers/95968-how-can-i-set-the-text-font-style-of-a-data-cursor-object
I do not know how and where to put the recommended commands. Would you please clarfity and give me more details?
Thank you.
Worth the wait: seven new online training courses and one new learning path were released with 25a, covering topics in Controls, Electrification, and Physical Modeling. This release also brings new functionality to support interactions across both MATLAB and Simulink within a single course, beginning with the new Controls courses below:
- Control System Design with MATLAB and Simulink (learning path – includes the 5 controls courses listed below)
- Control System Modeling Essentials
- Linearization of Nonlinear Systems
- Control System Analysis Techniques
- PID Control Techniques
- Classical Controller Design Techniques
- Battery State Estimation
- Motor Modeling with Simscape Electrical
Did you know that function double with string vector input significantly outperforms str2double with the same input:
x = rand(1,50000);
t = string(x);
tic; str2double(t); toc
tic; I1 = str2double(t); toc
tic; I2 = double(t); toc
isequal(I1,I2)
Recently I needed to parse numbers from text. I automatically tried to use str2double. However, profiling revealed that str2double was the main bottleneck in my code. Than I realized that there is a new note (since R2024a) in the documentation of str2double:
"Calling string and then double is recommended over str2double because it provides greater flexibility and allows vectorization. For additional information, see Alternative Functionality."
This just came out. @Michelle Hirsch spoke to Jousef Murad and answer his questions about the big change in the desktop in R2025a and explained what was going on behind the scene. Enjoy!
The Big MATLAB Update: Dark Mode, Cloud & the Future of Engineering - Michelle Hirsch
These got released last week and the process for using them on your local machine with MATLAB is very similar to how you use the local deepseek models as I demonstrated in my February blog post How to run local DeepSeek models and use them with MATLAB » The MATLAB Blog - MATLAB & Simulink
You need Ollama and the LLMs with MATLAB package installed (Details on how to do this in the blog post above). Then you run the following in your operating systems' command line
ollama pull gpt-oss:20b
Over to MATLAB and set up a chat session
>> chat = ollamaChat("gpt-oss:20b")
chat =
ollamaChat with properties:
ModelName: "gpt-oss:20b"
Endpoint: "127.0.0.1:11434"
TopK: Inf
MinP: 0
TailFreeSamplingZ: 1
Temperature: 1
TopP: 1
StopSequences: [0×0 string]
TimeOut: 120
SystemPrompt: []
ResponseFormat: "text"
FunctionNames: []
txt = generate(chat,"Who are you?")
txt =
"I’m ChatGPT – a conversational AI developed by OpenAI. My core is the GPT‑4 language model, which has been trained on a massive mix of text from books, websites, articles and other sources to understand and generate human‑like language. I don’t have feelings, consciousness, or a personal identity; I’m a tool that can help answer questions, brainstorm ideas, explain concepts, draft text, and more. My goal is to understand the context you give me and respond in a helpful, accurate and safe way. If there’s something specific you’d like to know or do, just let me know!"
This is the smaller of the two, new open models and it is bringing my aging desktop to its knees. My GPU is too small to do the work so I think everything is happening on the CPU and its slooooow. Will try on my Mac next
Let me know if you try this out!
Long before I joined MathWorks, I was a member of the academic Research Software Engineering (RSE) community where part of my mission was to introduce basic software engineering concepts to the research community. Things like version control, testing and even simply writing code instead of using only pointy-clicky GUIs before copying and pasting the results plot into a word document. I've seen things..........*shudders*
The RSE movement is still going very strong and I am elated that MathWorks is increasingly interacting with it. One example of such interaction is a video tutorial contributed by my colleauge @Mihaela Jarema to a comminity seminar series called 'A summer of Testing' It's linked to below
The video assumes you've never run a test before and gently guides you through the principles. Along the way you'll learn about some of MATLAB's superb testing capabilities. Things like
- Unit testing Framework
- Test Browser App
- Code Coverage
- Test Fixtures (Setup and teardown)
- Test driven devellopment
- Function argument validation
- CI/CD using GitHub actions
Go check out out.
Hey cody fellows :-) !
I recently created two problem groups, but as you can see I struggle to set their cover images :
What is weird given :
- I already did it successfully twice in the past for my previous groups ;
- If you take one problem specifically, Problem 60984. Mesh the icosahedron for instance, you can normally see the icon of the cover image in the top right hand corner, can't you ?
- I always manage to set cover images to my contributions (mostly in the filexchange).
I already tried several image formats, included .png 4/3 ratio, but still the cover images don't set.
Could you please help me to correctly set my cover images ?
Thank you.
Nicolas
Is there a hardware support package available for the MP series?
I just wanted here to share a link to some .gif animations I created over the years with Matlab :-)
I think gif animations are great supports for scientific diffusion.
Hello to all!
I would like to share with the Matlab and Simulink community this video about Neural Networks in Simulink.
This is a series of videos that use a multilayer perceptron implemented in Simulink as a case study. Why Simulink? Because it's a visual and intuitive modeling tool, you can see the forward propagation of this network and better understand the flow. The objective of this series is to show the implementation using Simulink for both simulation and Arduino, as well as its training using Matlab and Matlab with Deep Learning Toolbox, and a video of training with Python.
The video is in Spanish, but the Simulink model is available in English for the entire community; subtitles are also available.
The files are located in the first comment of each video. We hope you find it interesting and enjoyable. Best regards!
Here I share the link to the first video.
