13

Deeper models can have advantages (in certain cases) Most people will answer "yes" to your question, see e.g. Why are neural networks becoming deeper, but not wider? and Why do deep neural networks work well?. In fact, there are cases where deep neural networks have certain advantages compared to shallow ones. For example, see the following papers The ...


11

I think in this case, you'll probably want to use a genetic algorithm to generate a topology rather than working on your own. I personally like NEAT (NeuroEvolution of Augmenting Topologies). The original NEAT paper involves evolving weights for connections, but if you only want a topology, you can use a weighting algorithm instead. You can also mix ...


7

Have you read the book The Master Algorithm: by Pedro Domingos? He discusses the present day machine learning algorithms... Their strengths, weaknesses and applications... Deep Neural Network Genetic Algorithm Bayesian Network Support Vector Machine Inverse Deduction


6

Why has this merge proven beneficial? If you think about the shared Value/Policy network as consisting of a shared component (the Residual Network layers) with a Value and Policy component on top rather than Separation of Concerns it makes more sense. The underlying premise is that the shared part of the network (the ResNet) provides a high-level ...


5

The other answer mentions NEAT to generate network weights or topologies. The paper NeuroEvolution: The Importance of Transfer Function Evolution and Heterogeneous Networks, which also gives a short summary of neuroevolution techniques, provides an alternative approach to NEAT. It uses Cartesian Genetic Programming to evolve multiple activation functions.


4

This sounds like a problem that might be solvable with a LSTM-DQN approach, as described in Language Understanding for Text-based Games using Deep Reinforcement Learning by Narasimhan et al., 2015, and then extended to a domain very similar to your problem in Deep Reinforcement Learning for Syntactic Error Repair in Student Programs by Gupta et al., 2019. ...


4

A more efficient way would be creating a multi input model, with something like this: ___________ _____________ |__Image__| |Other input| _____|_____ _____|_____ |___CNN___| |__Dense___| _____|______ _____|______ |_Features1_| |_Features2_| __|_____|__ |__Merge___| _____|______ |___Dense__| ...


4

Although there is a strong element of "try and see" that has driven successful architectures, the drivers for what to try are often inspired by underlying theory or knowledge from other disciplines. Specifically for basic CNN, which led to AlexNet and many of the best image processing, the concept of using local receptive fields in layers was inspired by ...


4

As I understand, I think you wish to directly try out some deep learning stuff and things like library downloading, tools downloading, and managing all these really stop you from even starting to try out deep learning experiments.If this is what you asked for: Google Colab I think this is the best place for you. Anyone with a Google Drive account can sign ...


4

Deep learning is actually pretty useful (relative to other techniques) precisely when there is no simple mapping between input and output, and features from the raw input need to be aggregated and combined in complex ways by successive layers to form the output. As I pointed out in my answer to the AI SE decompilation question, there is recent DL research ...


4

I've read all the papers about PReLU, LeakyReLU (...) and all the claims how it improves this and that but the little dirty secret is: most of the time it doesn't matter at all and you can't go much wrong with ReLU - empirically proven. I've personally tried all of them in many different problems (from training small networks from scratch through changing ...


3

Estimating from an observation is a function, but "really counting" is a process. Feed-forward neural networks can learn arbitrary functions from training examples, but they cannot represent (and therefore cannot learn) processes. They can attempt to estimate the results of completing a process as a function, but that is not the same thing as ...


3

The Flatten layer is used for collapsing an ND tensor into a 1D tensor. In your case, the inputs appear to be $28\times28$ images, so Flatten will convert that into a tensor with shape $1\times768$. Note that no information is lost. Flatten layers are usually used where you have a convolutional layer with dimensions $N\times M \times C$ (where $N$,$M$ are ...


2

The networks in NEAT are still implicitly layered. There are neurons that need to be evaluated before other neurons can be evaluated and so this gives us our layers. If you don't know the structure of your network then you can use Kahn's algorithm to find an arbitrary (by arbitrary I just mean one of the possible partially ordered sets) ordering of the ...


2

I skimmed through your question and understood that the state/action space is finite, so in this case, RL would be a good option for storage. The most basic RL technique will keep track of a matrix Q ∈ ℝs×a, where s is number of possible states, and a is number of possible actions. In addition to a small overhead of agent's parameters: &...


2

The input to the neural network is a $19 × 19 × 17$ image stack comprising $17$ binary feature planes. $8$ feature planes $X_t$ consist of binary values indicating the presence of the current player’s stones ($X^i_t = 1$ if intersection $i$ contains a stone of the player’s colour at time-step $t$; $0$ if the intersection is empty, contains an opponent ...


2

There is indeed an investigation in progress, regarding this topic. A first publication from last march noted that modularity has been done, although not explicitly, since some time ago, but somehow training keeps being monolithic. This paper assess some primary questions about the matter and compares training times and performances on modular and heavily ...


2

In the perceptron design generally used in Artificial Neural Networks, we know precisely what a single neuron is capable of computing. It can compute a function $$f(x) = g(w^{\top} x),$$ where $x$ is a vector of inputs (may also be vector of activation levels in previous layer), $w$ is a vector of learned parameters, and $g$ is an activation function. We ...


2

The question is really broad---as stated by @thecomplexitytheorist---so difficult to give a meaningful answer. The following is about a clarification about the problem, and some directions. A model-based reflex agent is a blueprint describing the key components necessary to build that agent. It is an abstract architecture to guide the creation of concrete ...


2

Deeper networks have more learning capacity in the sense that they can fit to more complex data. But at the same time, they are also more prone to overfitting the training data and therefore fails to generalize to the test set. Apart from overfitting, exploding/vanishing gradients is another problem which hampers convergence. This can be addressed by ...


2

My experience from a tactical standpoint is to start out with a smaller simple model first. Train the model and observe the training accuracy and validation loss and validation accuracy. My observation is that, to be a good model, your training accuracy should achieve a value of at least 95%. If it does not, then try to optimize some of the hyper-parameters. ...


2

First of all, there is no real 'intelligence' innate to artificial Neural Networks (NNs). All they do is trying to approximate a mathematical function with a certain degree of generalization (hopefully without learning a given dataset by heart, i.e. hopefully without overfitting). The more nodes (or neurons) you include into the network, the more complex a ...


2

There's a few reasons I can think of, though I have not read an explicit description of why it is done this way. It's likely that people just started doing it this way because it's most logical, and people who have attempted to try your method of having reduced connections have seen a performance hit and so no change was made. The first reason is that if you ...


2

This type of connections are called skip or residual connections. There are numerous works which employs this type of mechanism, for example: ResNet, SkipRNN. In addition here you can find a paper that empirically explores the skip connections for sequential tagging, or this one for speech enhancement.


2

A unified neural network model consists of one neural network as opposed to other models that rely on two or more neural networks. For example, from page two of the YOLO paper: 2. Unified Detection We unify the separate components of object detection into a single neural network. Our network uses features from the entire image to predict each bounding box. ...


1

Convolutional layers are added in order to extract features from the image (like edges, corners, textures). After extracting those features, you feed them to a fully connected neural network to get the prediction. Let's take an example, consider you want to classify the cat's image. But you decided to do this by only using the convolutional layer. So, you ...


1

It depends on whether the action is part of the input or output of a neural network estimating the Q-value(state, action). The network on the left has the state as input and outputs one scalar value for each of the categorical actions. It has the advantage of being easy to setup and only needs one network evaluation to predict the Q-value for all actions. ...


1

I'm not aware of a direct way for finding the best NN architecture for a given task, but the recommended way, as far as I know, is to devise a network that can overfit the training data, and then apply regularization on top of it. That way, you can be almost sure you're not underfitting/underperforming due to network capacity.


Only top voted, non community-wiki answers of a minimum length are eligible