Skip to content

Implementations of various deep learning architectures + extra theoretical information

Notifications You must be signed in to change notification settings

ada-k/DeepLearningArchitectures

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

25 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DeepLearningArchitectures

Implementations of various deep learning architectures + extra theoretical information

Feature Extraction Local connectivity:

  1. Apply a feature extraction filter == convolution layer.
  2. Add bias to the filter function.
  3. Activate with a non-linear function e.g ReLU for thresholding.
  4. Pooling: dimesnionality reduction of generated feature maps from 1&2 above. Still maintains spatial invariance.

Classification:

  1. Apply a fully connected layer that uses the high-level features from 1-2-3 to perform classification.

Other potential applications; to achieve, you only modify the fully-connected layer bit == step1 under classification.

  • Segmentation.
  • Object Detection.
  • Regression.
  • Probabilistic Control.
  • Autonomous navigation.

Applications:

  • Density Estimation.
  • Sample Generation.
  • Input data X is encoded(compressed, self-encoded, auto encoded) to a lower dimensional latent space Z then model(decoder network) learns from the Z to reconstruct x as x hat.
  • Loss function == sq of difference between pixel by pixel difference between x and x hat.
  • VAEs introduces a stochasti process/randomness/probability aspect by calculating mean and SD from which the latent space features Z are sampled from.
  • VAEs loss = reconstruction loss + regularisation term(enforcess continuity and completeness).
  • Regularisation(D) = distance between 2 distributions (techniques like KL-divergence can be used to qunatify) image (prior is introduced)
  • To counter the challenge of backpropagation cause of stochasticty, reparametrisation is introduced == calculating a fixed vector of means and SDs thus driving away the probabilistic nature away from means and SDs of z.
  • Generators starts from noise(z==latent space) then imitates input data based on this.
  • Discriminator tries to identify real data from fakes created by the generator.(minimises probability of fake)
  • The process is iterative till the discriminator produces the highest probability that the generated data is real.
  • G == global optimum = produce the true data distribution == minmax
  • output at time t = a function of the input at t and past memory at time t-1.

Steps:

  1. Initialise the weight matrices and the hidden state to 0.
  2. Defining the core function: the forward pass through:
  3. Update of the hidden state(input + previous state).
  4. Output computation to generate output and new hidden state.

Design Criteria:

  • Handle sequences of variable length.
  • Track longterm dependencies.
  • Maintain info about order.
  • Share parameters across the sequence.

3.2 LSTM

Concepts:

  1. Maintains cell state.
  2. Has gates that control info flow: eliminates irrelevantt, keeps relevant.
  3. Backpropagation through time with partially uninterrupted gradient flow.

...

...

...

Layers:

  • Input layer - Receives raw input data and passes the info to follow up layers - hidden. No computation occurs here.
  • Hidden layer - An abstraction of the fully formed network. Performs computation on features from input layer(or another hidden layer) and passes result to output laye r(or another hidden layer). PS: They all use the same activation function.
  • Output layer - Final layer of the network hilding final value/prediction/output of the network. Has a different activation function depending on the goal.

Feedforward Network

Connectivity: Forward flow of information: Input used to calculate intermediate function in the hiddent layer to generate an output.

  1. Multiplication of input with neuron weights.
  2. Addition of bias.
  3. Passing output through an activation function.
  4. Output layer.

Backpropagation

  • Repeated adjusting of network weights and bias to minimise the cost/loss function based on the previous epoch.
  • The level of adjustment is determined by the gradients of the cost function with respect to those parameters.
  • Connectivity - calculating gradients of the loss/error function, then updating existing parameters in response to the gradients image

Weights and Bias

  • Weight - Learnable. Transforms input data within the hidden layers.
  • Bias - Learnable. Represents how off the predictions are from the expected values. == error term but not cost function.
  • Root purpose: not all data is linear. Performing a simple calculation of weights and bias would otherwise be a normal layered linear regression.
For Hidden Layers:
  1. Rectified Linear Activation (ReLU)
  • Function is linear for values > 0 but still non linear since all -ve values are tranformed to 0
# g(z) = max{0, z}
if input > 0:
	return input
else:
	return 0
  1. Logistic (Sigmoid)
  • Takes real values as input and outputs them in the range(0,1)
g(x) = 1.0 / (1.0 + e^-x)
  • Prone to problem of vanishing gradients.
  1. Hyperbolic Tangent (Tanh)
  • Takes real value inputs and outputs them in the range(-1, 1)
  • Takes same shape as a sigmoid.
g(x) = (e^x – e^-x) / (e^x + e^-x)
  • Prone to problem of vanishing gradients.
For Output Layers:
  1. Linear
  • == no activation/identity. Returns output as is.
  1. Logistic (Sigmoid)

  2. Softmax

  • Outputs a vector of values that sum up to 1.0. Like probabilities.
g(x) = e^x / sum(e^x)

Optimisation Techniques

...

...

...

...

...

...

...

...

...

...

Gradient Issues:

  • Vanishing gradients - During back propagation, gardient is calucalted w.r.t the params(weights and bias) with a goal of minimising the cost function. In the process, the gradients descennds downwards and often gets smaller till it hits a 0, -- at this point, the weights and bias remain almost unchanged resulting to the 0 chance of hitting a global optimum.

  • Vanishing Flags:

  1. Significant change in higher level params compared to slight/or no change in lower level params.
  2. Weights may become 0 during training.
  3. Models learns slowly and may stagnate during early iterations.
  • Example Cause:Certain activation functions, e.g sigmoid results in a shrink in variance between the input and output values s.t large inputs saturate close to 0.

  • Exploding gradients - In this case, the gradient ascends upwards resulting in huge weight and bias adjustments causing the gradient to diverge, hence explodes.

  • Exploding Flags:

  1. Exponential growth in model parameters.
  2. Model experienced avalanche training.
  3. Model weights may appear as NaN during training.
  • Example Cause: Initialisation params that result in large cost functions >> gardients can accumulate during an update resulting to large updates.

Convergence:

  • Global optimisation - It is the extreme of the objective function for the entire input space. Multimodal Optimisation is a case where more than 1 global optimas exist. Algos include: Genetic, Simulated Annealing, Particle Swarm Optimization.
  • Local optimisation - Maximisation or Minimisation of objective functions for a given region. If an objective function has a single local optima, then it is by default the global optima. Local search is the process of finding the local optima. Global optima can also be located by local search if its the basin under focus. Algos include: hill-climbing, Nelder-Mead, BFGS.
  • Techniques that modify learning algos by penalising weight matrices of nodes- so they can generalise better.
1. L2 and L1

The loss function of a network is extended by a regularisation term.
Reg term is multiplied by reg rate (alpha) before being added to the loss function.
Cost function = Loss (say, binary cross entropy) + Regularization term
With this, the gradient can be calculated with the updated loss function and use it to update the weights in back propagation.

image left=L1, right=L2
Source: An Introduction to Statistical Learning by Gareth James, Daniela Witten, Trevor Hastie, Robert Tibshirani

  • L2/Weight Decay/Ridge Regression - Euclidean norm of the weight matrices == sum over all squared weight values of a weight matrix.
    Forces weights to decay towards 0 but not 0 -- cause by adding a reg term, we introduce an additional subtraction from the current weights.

  • L1/Lasso Regression - Reg term is the absolute sum of weight values of a weight matrix.
    Forces the weights all the way towards 0.

PS: Smaller weight values reduces the impact of nodes in hidden layers thus achieving a simple model that doesn't model noise. The trade off is in the alpha value:

  • If alpha is too high, the model will be too simple and chances of underfitting are high.
  • If alpha is too low, the model will lbe too complex and we're back to the problem of overfitting.
2. Dropout
  • During training, a neuron gets turned off with some probability P.
  • Applied at every forward propagation and weight update step.
  • Preferred in large networks to utilise the randomness.

image Source: Journal of Machine Learning Research 15 (2014)

3. Data Augmentation

Increasing size of training data e.g by rotating, flipping, scaling or shifting an image.

4. Early Stopping

A cross-validation technique that sets aside a validation set, and stops the model training when the validation set accuracy is getting worse.

  • Text -
  • Images -
  • Videos -

About

Implementations of various deep learning architectures + extra theoretical information

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages