Leonardo Lucio Custode

Senior AI Researcher.

Opinions are my own.

[Leonardo Lucio Custode @LLCustode]

@udaysy @fchollet Even neural networks can be inspected, modified and understood up to a certain size. The problem is the scale of the software/networks that makes them blackboxes: when the effort of inspecting them becomes sufficiently large, you're likely going to treat them as a black boxes

[Leonardo Lucio Custode @LLCustode]

@trobuling @fchollet Depending on your position in the company, I think it may be your job to inform your employer (or anyone above you) about potential wrong hardware/infrastructure choices

[Leonardo Lucio Custode @LLCustode]

@EdSealing @DamienTeney @ziv_ravid Well, not necessarily, they might be very different paths. Humans have their own biases because of the evolutionary path that led us here. Going from 0 to human, and then from human to optimal might be more expensive than 0 to optimal (e.g. alphazero vs alphago)

[Leonardo Lucio Custode @LLCustode]

@DamienTeney @EdSealing @ziv_ravid Also, imo, sticking too much to recreating human-like AI will basically have little value. If you could replicate with 100% accuracy a human brain, it would have the same limitations as humans, which by definition will not enable any super-human behavior

[Leonardo Lucio Custode @LLCustode]

@karpathy Gemini + https://t.co/j7rtOaysmO is the way

[Leonardo Lucio Custode @LLCustode]

@NeuralRunner The special issue is open to both methodological contributions and to applications to any area where interpretability could be a great added value

[Leonardo Lucio Custode @LLCustode]

Working on Interpretable Reinforcement Learning? Submit your work to our Special Issue on Applied Soft Computing!

Submission: Dec 31, 2025
Final Decision: June 1, 2026

Details & submission: https://t.co/H4yJ7XoVvU

#reinforcementlearning #interpretability #xai

[Leonardo Lucio Custode @LLCustode]

@seth_quant Yes, and in many cases they can match the performance of non-interpretable methods! Looking forward to submissions that close the gap in many other areas where Interpretable RL is still behind non-interpretable methods!

[Leonardo Lucio Custode @LLCustode]

Organizers: @gih82 @GNadizar @EricMedvetTs Erica Salvato, and I

[Leonardo Lucio Custode @LLCustode]

@emollick Is it super-human in *all* the experiments though? I see that the results are not statistically significant in various plots

[Leonardo Lucio Custode @LLCustode]

@vineettiruvadi @_jasonwei Can you list some examples?

[Leonardo Lucio Custode @LLCustode]

@skinnnnnnnner @fchollet Can you elaborate on this please?

[Mátyás Vincze @vinczematyas_]

🎉 Our work "SMOSE: Sparse Mixture of Shallow Experts for Interpretable Reinforcement Learning in Continuous Control Tasks" has been accepted by AAAI 2025! #AAAI2025

paper: https://t.co/6nBahjMLmn
with @L_Ferrarotti, @brulepri, @gih82

🧵1/6

[Giovanni Iacca @gih82]

New paper out! Fast-Inf: Ultra-Fast Embedded Intelligence on the Batteryless Edge https://t.co/oxlQm0cHgt #tinyml #machinelearning #artificialintelligence

[Leonardo Lucio Custode @LLCustode]

@VictorTaelin Would you say that this is related to the architecture or to the way the architecture is currently trained (i.e., data-driven)?

[Leonardo Lucio Custode @LLCustode]

Happy to share that we won the Interpretable Control Competition @ #GECCO24 !

Congrats @vinczematyas_ @L_Ferrarotti @gih82 @brulepri

[Giovanni Iacca @gih82]

Our @GeccoConf papers are finally out!
Multi-Objective Evolutionary Hindsight Experience Replay for Robotics https://t.co/NFFOmBrHx8
Decentralized Federated NeuroEvolution of Heterogeneous Networks https://t.co/tKyxdGkLHR
Neuron-centric Hebbian Learning
https://t.co/UbsTHxQcIm

[Leonardo Lucio Custode @LLCustode]

@jsuarez If you look at the search space, in the discrete case you have to pick a value from N possibilities (N is the number of actions), while in the continuous case you have to choose N values from R^N or [-1, 1]^N in practice, which makes the search much larger and harder to optimize

[Leonardo Lucio Custode @LLCustode]

@eyeslasho What's the y axis?

[Leonardo Lucio Custode @LLCustode]

@adellacioppa @gih82 @CunegattiElia @ayaman00 @facaraff In "Comparing Large Language Models and Grammatical Evolution for Code Generation" (w/ @ChiaraRambaldi , Marco Roveri, @gih82), we compare LLMs and evolutionary techniques in code generation tasks.

5/5

[Leonardo Lucio Custode @LLCustode]

Excited to share that we got four papers accepted @ GECCO 2024!

Here's a brief recap of our papers!

[Leonardo Lucio Custode @LLCustode]

@adellacioppa @gih82 @CunegattiElia In "An investigation on the use of Large Language Models for hyperparameter tuning in Evolutionary Algorithms" (w/ @ayaman00, @facaraff , @gih82 ), we investigate the capabilities of LLMs in tuning hyperparameters in evolution strategies.

4/5