arXiv subjects

Alessandro Marchetti

Publications and source records attributed to Alessandro Marchetti.

4 recordsLinked to original sources

arXiv2023

Increasing biases can be more efficient than increasing weights

We introduce a novel computational unit for neural networks that features multiple biases, challenging the traditional perceptron structure. This unit emphasizes the importance of preserving uncorrupted information as it is passed from one unit to the next, applying activation functions later in the process with specialized biases for each unit. Through both empirical and theoretical analyses, we show that by focusing on increasing biases rather than weights, there is potential for significant enhancement in a neural network model's performance. This approach offers an alternative perspective on optimizing information flow within neural networks. See source code at https://github.com/CuriosAI/dac-dev.

cs.NE↗

arXiv2022

Score vs. Winrate in Score-Based Games: which Reward for Reinforcement Learning?

In the last years, the DeepMind algorithm AlphaZero has become the state of the art to efficiently tackle perfect information two-player zero-sum games with a win/lose outcome. However, when the win/lose outcome is decided by a final score difference, AlphaZero may play score-suboptimal moves because all winning final positions are equivalent from the win/lose outcome perspective. This can be an issue, for instance when used for teaching, or when trying to understand whether there is a better move. Moreover, there is the theoretical quest for the perfect game. A naive approach would be training an AlphaZero-like agent to predict score differences instead of win/lose outcomes. Since the game of Go is deterministic, this should as well produce an outcome-optimal play. However, it is a folklore belief that "this does not work". In this paper, we first provide empirical evidence for this belief. We then give a theoretical interpretation of this suboptimality in general perfect information two-player zero-sum game where the complexity of a game like Go is replaced by the randomness of the environment. We show that an outcome-optimal policy has a different preference for uncertainty when it is winning or losing. In particular, when in a losing state, an outcome-optimal agent chooses actions leading to a higher score variance. We then posit that when approximation is involved, a deterministic game behaves like a nondeterministic game, where the score variance is modeled by how uncertain the position is. We validate this hypothesis in AlphaZero-like software with a human expert.

cs.AI↗

arXiv2004

Performances of the Italian seismic network, 1985-2002: the hidden thing

Seismic data users and people managing a sesimic network are both interested in the potentiality of the data, with the difference that the former look at stability, the second at improvements. In this work we measure the performances of the Italian Telemetered Seismic Network in 1985-2002 by defining basic significant parameters and studying their evolution during the years. Then, we deal with the geological methods used to characterise or to plan a seismic station deployment in a few cases. Last, we define the gain of the network as the percentage of located earthquakes with respect to the total recorded earthquakes. By analysing the distribution of non-located ("missed") earthquakes, we suggest possible actions to take in order to increase the gain. Results show that completeness magnitude is 2.4 in the average over the analysed period, and it can be as low as 2.2 when we consider non-located earthquakes as well. Parameters such as the distance between an earthquake and the closest station, and the RMS location decrease with time, reflecting improvements in the location quality. Methods for geologic and seismological characterisation of a possible station site also proved to be effective. Finally, we represent the number of missed earthquakes at each station, showing that nine stations control more that 50% of all missed earthquakes, and suggesting areas in Italy where the network might be easily improved.

physics.geo-ph↗

arXiv2003

Holographic three-point functions: one step beyond the tradition

Within the program of holographic renormalization, we discuss the computation of three-point correlation functions along RG flows. We illustrate the procedure in two simple cases. In an RG flow to the Coulomb branch of N=4 SYM theory we derive a compact and finite expression for the three-point function of lowest CPO's dual to inert scalars. In the GPPZ flow, that captures some features of N=1 SYM theory, we compute the three-point function with insertion of two inert scalars and one active scalar that mixes with the stress tensor. By amputating the external legs at the mass poles we extract the trilinear coupling of the corresponding superglueballs. Finally we outline the procedure for computing three-point functions with insertions of the stress tensor as well as of (broken) R-symmetry currents.

hep-th↗