abtexp/jbrain

Neural network in javascript

Jbrain

Machine Learning In JavaScript

// For latest version (unstable) clone the dev branch.

Current Progress :

• Implementing A Generic API For Creating Different Types Of Neural Networks
• Implementing Different Cost And Activation Functions
• Implementing Different Optimzation Methods
• Implementing SVM
• Implementing K-means
• Activation Functions : relu, softmax, sigmoid, tanh
• Cost Functions : cross entropy, quadratic cost
• Optimizers : Gradient Descent (batch, mini batch and stochastic), RMSProp, Adam
• Network Types : Feed Forward

Future Work :

• Fixing All Bugs
• Convolutional and reccurent networks
• Weight Initializations
• Regularization Techniques
• Faster API Performance
• GPU Support

@abtExp

--Any help or suggestion is appreciated.

jbrain's Issues

weights are undefied ...

Probably an error due to vecto's arrange method. ❗️

https://github.com/abtExp/Jbrain/projects/2#card-4715002

Use vecto.Ndarray for operations.

Use Ndarray as the standard for all the inputs as well as operations.

part_err calculation err due to uneven size

The arg1 and 2 for product function when training the network are uneven in size.
`var partErr = product(this.weights[i].array,ele);
console.log("partErr = " + partErr + "z = " + this.Z[ip_num][i-1]);
let err = product(partErr,this.Z[ip_num][i-1]);

/* error : del contains only 1 element at the stage for calculating the error for the second last layer thus del[i+1] i.e., del[2] will be undefined */

del[i] = err;`

fix it.

Testing for the back and forward prop

writing unit tests in jest for testing props

the activation array is NaN

try to solve the NaN error

fix the constructor methods for new network creations

There's an error calculating delta for hidden layers

Have to implement transpose method in vecto as the delta for hidden layer is (w(i+1)^T.del(i+1))o sig_ (i);

error in reassigning the values to nw and nb, flat array unassigned

the flat array for nw and nb remain unchanged i.e., assigned as zeros.

Optimize the optimizer code

The optimizer_class has to provide a more generic and faster api for better performance.

Implement The CNN, SVM and KNN

Just've to write them, planning is done. Need to optimize the optimizer code first.

make and check the grads for all cost funcs and activ funcs

The cost funcs and activ funcs are to be tested and the grad parts are to be written
#12

Use keras style api for creating a model. Thus different layer types can be implemented without confusion.

possible miscalculation of the delta for hidden layers

the delta for ith hidden layer must be =
del(i) = ((w(i+1).del(i+1)).sigma_dash(i))

instead done
del(i) = ((w(i+1).del(i+1)).z(i))

also when fixed the formula, uneven size error emerged.