A naukri.com initiative
Neural Networks News
Hackernoon
2d
90
Image Credit: Hackernoon
On the Interpretability of Part-Prototype Based Classifiers: A Human Centric Analysis
- Part-prototype networks have recently become Methods of interest as interpretable alternative to many black box classifiers.
- The interpretability of these methods from the perspective of humans has not been sufficiently explored so far.
- A framework that evaluates the interpretability of part-prototype based models from a human perspective has been proposed in this work.
- The proposed framework consists of three actionable metrics and experiments focusing on prototype interpretability, prototype-query similarity and interpretability of decision-making process.
- Experiments were performed using Amazon Mechanical Turk, which showed the capability of the framework in assessing the interpretability of various part-prototype-based models.
- The experiments performed in this work are the most comprehensive work on evaluating part-prototype-based methods in a unified framework.
- Prototype-based classifiers are a category of inherently interpretable methods that uses prototypical examples to make their decisions.
- Newer methods have been proposed that combine the power of neural networks and the decision-making process of a prototype based classifier to create prototypical neural nets.
- The interpretability of a part-prototype system requires the interpretability of the prototype, the similarity of the prototype to the corresponding region in the query sample and the interpretability of the decision-making process.
- This work proposes three actionable metrics and experiments for evaluating the interpretability of part-prototype-based classifiers.
Read Full Article
5 Likes
Medium
3d
173
Image Credit: Medium
A Beginners Guide to PyTorch
- PyTorch is an open-source library for building neural networks.
- PyTorch is more Python-friendly and intuitive compared to TensorFlow.
- Basic operations such as tensor initialization, indexing, and arithmetic are explained.
- A basic PyTorch neural network is built, and training using DataLoader, loss function, and optimizer is covered.
Read Full Article
10 Likes
Medium
4d
387
Image Credit: Medium
LLMs don’t even do ‘approximate retrieval’ — embarrassingly, they try to recall some ‘similars’
- Counterfactual task is an ideal task to test LLMs by exposing them to data that is unlikely to have been seen during training.
- Disrupting the training 'template' of LLMs results in a drop in performance to near-random choice.
- Experiments conducted on GPT 4o reveal that LLMs are essentially massive fuzzy hashtables and do not understand language or reasoning.
- A purely behavioristic, associative, statistical paradigm cannot explain cognition, language, reasoning, understanding, and the mind.
Read Full Article
23 Likes
Medium
4d
313
Image Credit: Medium
Basic concepts and terms on how machine learns — Neural Networks
- The Neural Network works in three layers: input layer, hidden layer and output layer. Input layer includes the input data, weights, and biases of each input, defined by the machine. Hidden layer includes an activation function like ReLu, Sigmoid, Hyperbolic Tangent, and Softmax, which are defined by humans depending on the problem to solve.
- After getting the output from the activation function, we compare it with the actual value. We use loss or cost function for comparing predicted and actual values. Loss function finds the errors for single training and Cost function calculates errors for the entire training set.
- Loss and Cost functions are used differently for classification(as Binary Classification) and regression problems(MSE). The difference between the predicted and actual value is determined by these functions.
- To reduce the difference, the machine goes back and adjusts the weights, a process called backpropagation that uses gradient decent and learning rate. Gradient Descent finds the bottom through the angle of the slope, and learning rate determines the rate of learning speed.
- Hyperparameters that are set by humans are Learning Rate and Epochs. If the learning rate is too small, the training might take a long time. If it's too high, the model might not converge to the expected level of accuracy. Epochs determine how many times the learning iterates.
- Product Managers need to understand the essential components of Neural Networks, such as input, hidden, and output layers, Activation functions, and iterative nature of learning. This knowledge enables product managers to make informed decisions on feature implementations and collaborate seamlessly with Data Science and Machine Learning teams to enhance product outcomes.
Read Full Article
18 Likes
Towards Data Science
5d
195
Image Credit: Towards Data Science
Understanding Abstractions in Neural Networks
- Abstraction is a concept used in mathematics, programming, and daily life to throw away irrelevant details and focus on crucial assumptions. In neural networks, abstraction is implemented through non-injective mappings. By abstracting away irrelevant details, neural networks create invariant features and cluster same-class samples together, while keeping between-class samples separated. Abstraction decreases entropy while preserving mutual information, and is crucial for neural networks to generalize well in the input space. Abstraction is everywhere and a core element of cognition, allowing us to reduce detailed world into manageable pieces and learn anything general.
- Abstraction, in essence, is throwing away irrelevant details, focusing only on crucial assumptions.
- In neural networks, abstraction is implemented through non-injective mappings, ignoring differences in parts of the input space.
- Neural networks abstract away irrelevant details to create invariant features, cluster same-class samples, and keep between-class samples separated.
- Abstraction decreases entropy while preserving mutual information, and is crucial for neural networks to generalize well in the input space.
- Abstraction is a core element of cognition, allowing people to reduce detailed world into manageable pieces and learn anything general.
- Abstraction is implemented computationally in the brain through hierarchical processing, similar to artificial neural networks.
- The exact meaning of abstraction in neural networks is often unclear. Shifting the attention to the information ignored in the process can provide better insights into how it produces increasingly abstract and invariant features.
- Abstraction is crucial to understand how neural networks generalize in the input space and is a central role in cognition.
- Abstraction is everywhere and is a core element of cognition.
Read Full Article
11 Likes
Medium
5d
277
Image Credit: Medium
A Comparison of Neural Network Attention Mechanisms in Obfuscated JavaScript Detection
- New research into neural networks has brought more advanced attention mechanisms, allowing for better performance in language processing applications.
- Research aimed at detecting obfuscated JavaScript files used different artificial intelligence techniques from static to dynamic code analysis methods.
- The study designed different neural network architectures to test which would perform better in detecting obsufcated JavaScript, using attention mechanisms as a variable.
- A comparison between architectures revealed that LSTM and transformer models showed no significant difference in accuracy metrics but transformers demonstrated superior training and inference throughput.
- The choice of architecture depends on the computation resources available, as transformers suit environments with ample memory and prioritize inference speed, while LSTMs are preferable in resource-constrained settings.
- The bidirectional LSTM model exhibited strong performance, capturing long-range dependencies effectively and achieving high accuracy without overfitting.
- The dataset used was a subset of ETH Zurich’s JavaScript dataset containing a broad range of different use cases and the most commonly used obfuscation tools were used to create obfuscated material from the unbfuscated data.
- The study finds that detecting obfuscated JavaScript files requires attention mechanisms.
- Overall, LSTM and transformer models can be considered equivalent in terms of accuracy when classifying obfuscated scripts.
- Both attention mechanisms are suitable for this classification problem and the advisable architecture simply depends on the available computational resources.
Read Full Article
16 Likes
Medium
5d
160
Image Credit: Medium
Mastering Neural Networks: From Basics to Advanced Applications
- The journey of training a neural network begins with model setup.
- Activation functions, such as ReLU and softmax, significantly affect neural network performance.
- Integrating loss computation directly with logits improves numerical stability.
- Adam optimization algorithm facilitates faster convergence and more robust performance.
Read Full Article
9 Likes
Medium
6d
111
Image Credit: Medium
Training a single perceptron from scratch
- Activation function is the function which takes an input (here it takes a) and returns values within a range.
- Weights are the strengths which tells us which input is more important for predicting the output.
- Backpropagation is the algorithm used to update the weights and biases of a neural network going back through the neural network.
- The role of bias is to shift the result to a bit up or down to get a non-zero value.
Read Full Article
6 Likes
Hackernoon
6d
193
Image Credit: Hackernoon
Neural Network for Valuing Bitcoin: Conclusion, Acknowledgments, and Funding
- This study proposed a novel pricing framework for valuing bitcoin call options using a bivariate Merton jump-diffusion model to describe the behavior of bitcoin prices and transaction volumes affecting the price dynamics.
- An extended Black-Scholes equation was derived to obtain a trial solution for the associated Black-Scholes PDE for bitcoin call options with European features.
- An artificial neural network was introduced in solving the unconstrained optimization problem resulting from the trial solution to obtain efficient valuation of the financial derivative.
- The study compared the numerical results of valuing bitcoin call options using the normal Black-Scholes model and the Merton jump-diffusion model, and observed that the latter resulted in a more efficient valuation process.
- One of the limitations of this research lies in finding the optimal neural network configuration that ensures fair pricing of bitcoin options using the two proposed models. This optimal feature will be incorporated to avoid over-pricing or under-pricing of these option values, and future research will focus on this.
- The approach used in the study may be prone to artificial Google searches affecting sentiment-based data analysis and decisions; future work will examine the correlation between search-based data and tweets data to improve the valuation process.
- The study focused exclusively on jump-diffusion models for cryptocurrency pricing, but suggested expanding the set of stochastic processes considered to provide a more thorough treatment of cryptocurrency dynamics, using processes like variance gamma, normal inverse Gaussian, and generalized hyperbolic Levy motions.
- Incorporating alternatives like Levy processes could potentially improve model fitting, and evaluating a range of stochastic models could ultimately enhance financial engineering techniques tailored specifically to cryptocurrencies.
- The research received no external funding, and the second author thanks the Research Centre of AIMS-Cameroon for hosting him during the preparation of the manuscript.
Read Full Article
11 Likes
Hackernoon
6d
82
Image Credit: Hackernoon
Neural Network for Valuing Bitcoin: Availability of Data, Code and Materials, Contributions
- This paper discusses the availability of data, code, and materials for a neural network model used for valuing Bitcoin.
- The authors of the paper are Edson Pindza, Jules Clement Mba, Sutene Mwambi, and Nneka Umeorah.
- The corresponding author can be contacted for the availability of data, code, and materials.
- All authors contributed equally to the paper, and there is no conflict of interest.
Read Full Article
4 Likes
Hackernoon
6d
321
Image Credit: Hackernoon
Neural Network for Valuing Bitcoin: Numerical Results, Implementation and Discussion
- This paper discusses the use of a neural network approach to estimate the European call options for bitcoin based on numerical results and parameter approximations. The empirical analysis involves the use of Bitcoin historical closing prices and Google trend data and the descriptive statistics for this data set are presented as well as the parameters for numerical computation and parameter estimation. We tested two models: Model I and Model II. For Model I, the comparison of two optimizations, namely SGD and Adam, are made using pre-defined parameters, while for Model II, we set up and compare two artificial activation functions and two optimizers. The results of experiments suggest Model II with Adam optimizer and ReLU activation function as the optimal model for calculating Bitcoin call option. Finally, the empirical validation of the model using equity options datasets demonstrates the model's ability to effectively price options for securities with dynamics including frequent jumps and volatility clustering.
- The paper acknowledges that emerging cryptocurrency market features violate the standard market efficiency and no-arbitrage assumptions in developing the jump-diffusion model framework. Relaxing the efficiency and no-arbitrage assumptions is an important area of research, however, in light of limited cryptocurrency market history, as the market matures, sophisticated models can evolve.
Read Full Article
19 Likes
Hackernoon
6d
345
Image Credit: Hackernoon
Neural Network for Valuing Bitcoin: Methodology
- This study proposes a bivariate jump-diffusion model with a sentiment indicator for pricing and obtaining derivatives of Bitcoin futures contracts.
- The proposed model enables investors to calculate the price of call and put options on Bitcoin futures contracts.
- The preliminary pricing and derivatives valuation framework proposed can be built upon as the market evolves.
- The proposed bivariate jump-diffusion model and sentiment indicator capture essential dynamics but may require adjustments as knowledge develops.
- The proposed neural network approach serves as an alternative way to numerically solve the pricing PDE without discretizing the domain.
- Jump-diffusion models are continuous-time stochastic processes that reproduce stylized facts observed in asset price dynamics, such as mean-reversion and jumps.
- Levy processes appear as an alternative for capturing large deviations in asset prices.
- Levy processes allow sample paths with frequent discontinuities, enabling them to generate heavy-tailed distributions.
- Finite difference methods and Monte Carlo simulation are the standard techniques used to approximate the solution to extended Black-Scholes PDE.
- The proposed model provides an essential first step in applying financial engineering techniques to the cryptocurrency domain.
Read Full Article
20 Likes
Hackernoon
6d
366
Image Credit: Hackernoon
Neural Network for valuing Bitcoin: References
- This paper presents a methodology based on artificial neural networks to approximate the partial differential equation (PDE) often derived for option pricing.
- The neural network approach is trained to find a numerical solution to the PDE, which is then used to analyze the real-time valuation of Bitcoin.
- The numerical results show that the proposed neural network method can provide accurate valuations of Bitcoin with a small error margin compared to Monte Carlo simulations.
- The paper also discusses the previous research studies that applied neural networks to solve PDEs in finance.
- The study shows that the proposed method can accurately value Bitcoin options and other derivatives while overcoming the curse of dimensionality.
- It also establishes the effectiveness of neural networks and their applicability in finance.
- The paper also cites previous research studies that have used neural networks to predict Bitcoin prices based on information available on online forums, news articles, and trends on Google and Wikipedia.
- It lists studies that have focused on estimating Bitcoin volatility, detecting jump risk, jump-diffusion model for Bitcoin option pricing, and sentiment and Bitcoin volatility.
- The authors of the research propose a neural network-based approach to finding an accurate numerical solution to partial differential equations, which would enable setting accurate values of Bitcoin.
- The proposed method can also be applied to valuing other derivatives and is shown to be effective in overcoming the curse of dimensionality.
Read Full Article
22 Likes
Hackernoon
6d
329
Image Credit: Hackernoon
Neural Network for Valuing Bitcoin: Abstract and Introduction
- This study considers a bivariate jump-diffusion model to describe Bitcoin price dynamics and the number of Google searches affecting the price, representing a sentiment indicator.
- A closed formula for the Bitcoin price and Black-Scholes equation for Bitcoin options valuation were derived.
- The model uses artificial neural networks to solve the corresponding Bitcoin option partial differential equation for pricing process.
- The prediction performance and model validation using various high-volatile stocks were assessed.
- Bitcoin price is found to be influenced by sentiment about the Bitcoin system and transaction popularity among other factors.
- Jump-diffusion model with a sentiment indicator variable provides a foundation for pricing and derivatives valuation in cryptocurrency markets.
- The use of search-based data as sentiment indices has the potential to reveal the underlying beliefs of populations directly.
- While traditional financial theory assumes normal distributions, jumps account for extreme price fluctuations beyond what continuous diffusion alone predicts in cryptocurrency markets.
- The proposed model offers an initial step towards financial engineering in the cryptocurrency space.
- The study presents a new research approach that can generate greater insights into the factors affecting Bitcoin price dynamics.
Read Full Article
19 Likes
Hackernoon
6d
102
Image Credit: Hackernoon
Neural Network for Valuing Bitcoin: Neural Network Methodology
- Neural Network (NN) algorithm is used for solving problems relating to dimensionality reduction, visualization of data, clustering and classification issues and is also used in regression in solving regression problems since they do not require prior knowledge about the distribution of the data. Recently, the NN has become an indispensable tool for learning the solutions of differential equations and in solving equations without analytical solutions.
- Bitcoin options with the corresponding Black-Scholes pricing PDE can be re-written as a non-homogeneous PDE, and standard theorems guarantee a classical smooth solution exists for the pricing PDE given the assumed dynamics. The continuity and linear growth conditions on the coefficient functions ensure they satisfy Lipschitz continuity.
- The neural network is introduced as an approximating function to solve the PDE and is utilized to approximate the price of the option numerically. By using the ANN in solving the PDE, the loss or cost function associated with the training is measured by how well the approximation function satisfies the differential operator, boundary conditions and the terminal condition of the option pricing PDE.
- The whole procedure for seeking a good parameter set for the neural network by minimizing the loss function using the gradient descent-based optimizers is called training and can be implemented in solving the PDE in equation (3.13). The learning rate employed to scale the intensity of the parameter updates during the gradient descent affects the pace at which the algorithm learns and whether or not the cost function is minimized.
- While the pricing PDE has a proven smooth solution under the model dynamics, obtaining the analytical form is intractable. Numerical methods must be used to approximate the solution. Neural networks provide a flexible parametric approach based on their universal approximation theoretical results. Future work should explore neural network training enhancements and theoretical guarantees to ensure robust solutions.
Read Full Article
6 Likes
For uninterrupted reading, download the app