Jeff Johnson

Mostafa Elhoushi, Jeff Johnson

We present any4, a learned 4-bit weight quantization solution for large language models (LLMs) providing arbitrary numeric representations without requiring pre-processing of weights or activations. any4 yields higher accuracy compared to other related 4-bit numeric representation types: int4, fp4 and nf4, as evaluated on a range of model sizes, generations and families (Llama 2, Llama 3, Mistral and Mixtral). While any4 does not require preprocessing of weights or activations, it is also competitive with orthogonal techniques that require such preprocessing (e.g., AWQ and GPTQ). We also experiment with any3 and any2 and show competitiveness at lower bits. Additionally, we show that we can calibrate using a single curated diverse sample rather than hundreds of samples from a dataset as done in most quantization approaches. We also open source tinygemm, a latency optimized GPU matrix multiplication library for LLMs, that implements any4 using a GPU-efficient lookup table strategy along with other common quantization methods. We open source our code at https://github.com/facebookresearch/any4 .

Kevin Moran, David N. Palacio, Carlos Bernal-Cárdenas et al.

Traceability is a fundamental component of the modern software development process that helps to ensure properly functioning, secure programs. Due to the high cost of manually establishing trace links, researchers have developed automated approaches that draw relationships between pairs of textual software artifacts using similarity measures. However, the effectiveness of such techniques are often limited as they only utilize a single measure of artifact similarity and cannot simultaneously model (implicit and explicit) relationships across groups of diverse development artifacts. In this paper, we illustrate how these limitations can be overcome through the use of a tailored probabilistic model. To this end, we design and implement a HierarchiCal PrObabilistic Model for SoftwarE Traceability (Comet) that is able to infer candidate trace links. Comet is capable of modeling relationships between artifacts by combining the complementary observational prowess of multiple measures of textual similarity. Additionally, our model can holistically incorporate information from a diverse set of sources, including developer feedback and transitive (often implicit) relationships among groups of software artifacts, to improve inference accuracy. We conduct a comprehensive empirical evaluation of Comet that illustrates an improvement over a set of optimally configured baselines of $\approx$14% in the best case and $\approx$5% across all subjects in terms of average precision. The comparative effectiveness of Comet in practice, where optimal configuration is typically not possible, is likely to be higher. Finally, we illustrate Comets potential for practical applicability in a survey with developers from Cisco Systems who used a prototype Comet Jenkins plugin.

Cristian Jimenez-Romero, David Sousa-Rodrigues, Jeffrey H. Johnson

This study explores the design and control of the behaviour of agents and robots using simple circuits of spiking neurons and Spike Timing Dependent Plasticity (STDP) as a mechanism of associative and unsupervised learning. Based on a "reward and punishment" classical conditioning, it is demonstrated that these robots learnt to identify and avoid obstacles as well as to identify and look for rewarding stimuli. Using the simulation and programming environment NetLogo, a software engine for the Integrate and Fire model was developed, which allowed us to monitor in discrete time steps the dynamics of each single neuron, synapse and spike in the proposed neural networks. These spiking neural networks (SNN) served as simple brains for the experimental robots. The Lego Mindstorms robot kit was used for the embodiment of the simulated agents. In this paper the topological building blocks are presented as well as the neural parameters required to reproduce the experiments. This paper summarizes the resulting behaviour as well as the observed dynamics of the neural circuits. The Internet-link to the NetLogo code is included in the annex.

Cristian Jimenez-Romero, David Sousa-Rodrigues, Jeffrey H. Johnson et al.

A model of an Ant System where ants are controlled by a spiking neural circuit and a second order pheromone mechanism in a foraging task is presented. A neural circuit is trained for individual ants and subsequently the ants are exposed to a virtual environment where a swarm of ants performed a resource foraging task. The model comprises an associative and unsupervised learning strategy for the neural circuit of the ant. The neural circuit adapts to the environment by means of classical conditioning. The initially unknown environment includes different types of stimuli representing food and obstacles which, when they come in direct contact with the ant, elicit a reflex response in the motor neural system of the ant: moving towards or away from the source of the stimulus. The ants are released on a landscape with multiple food sources where one ant alone would have difficulty harvesting the landscape to maximum efficiency. The introduction of a double pheromone mechanism yields better results than traditional ant colony optimization strategies. Traditional ant systems include mainly a positive reinforcement pheromone. This approach uses a second pheromone that acts as a marker for forbidden paths (negative feedback). This blockade is not permanent and is controlled by the evaporation rate of the pheromones. The combined action of both pheromones acts as a collective stigmergic memory of the swarm, which reduces the search space of the problem. This paper explores how the adaptation and learning abilities observed in biologically inspired cognitive architectures is synergistically enhanced by swarm optimization strategies. The model portraits two forms of artificial intelligent behaviour: at the individual level the spiking neural network is the main controller and at the collective level the pheromone distribution is a map towards the solution emerged by the colony.

Matteo Rucco, David M. S. Rodrigues, Emanuela Merelli et al.

This work introduces an integrative approach based on Q-analysis with machine learning. The new approach, called Neural Hypernetwork, has been applied to a case study of pulmonary embolism diagnosis. The objective of the application of neural hyper-network to pulmonary embolism (PE) is to improve diagnose for reducing the number of CT-angiography needed. Hypernetworks, based on topological simplicial complex, generalize the concept of two-relation to many-body relation. Furthermore, Hypernetworks provide a significant generalization of network theory, enabling the integration of relational structure, logic and analytic dynamics. Another important results is that Q-analysis stays close to the data, while other approaches manipulate data, projecting them into metric spaces or applying some filtering functions to highlight the intrinsic relations. A pulmonary embolism (PE) is a blockage of the main artery of the lung or one of its branches, frequently fatal. Our study uses data on 28 diagnostic features of 1,427 people considered to be at risk of PE. The resulting neural hypernetwork correctly recognized 94% of those developing a PE. This is better than previous results that have been obtained with other methods (statistical selection of features, partial least squares regression, topological data analysis in a metric space).