Emergence of communication among agents based on deep learning with predictive processing

Komunikacija med agenti je nujna za učinkovito medsebojno sodelovanje, saj je verjetno prav razvoj jezika ključen za človeški napredek. To delo poskuša doseči samoorganizacijo komunikacijskega sistema med dvema agentoma v računalniški simulaciji na način, ki ima teoretično zasnovo v nevroznanosti. Jezik se smatra kot komunikacijski protokol, ki spaja udeležence v pogovoru v večji sistem, v katerem se optimalno delovanje samoorganizira preko optimizacije posameznikov. Agenta, ki sta udeležena v komunikaciji, sta zasnovana po principu napovednega procesiranja, pri katerem agenti za razumevanje sveta zmanjšujejo napako med napovedjo o prihodnjih stanjih sveta in sebe ter dejanskimi zaznanimi signali. S tem zmanjšujejo presenečenje na vhodnih podatkih in posledično izboljšujejo svoj generativni model sveta. Postavljena je preprosta simulacija z dvema agentoma, ki vidita isto sliko in imata zmožnost tvorjenja sporočil o njej. Agenta sta modelirana kot zaznavna in generativna nevronska mreža na vsaki od obeh modalnosti – vidni in besedni – s skupno notranjo predstavitvijo konceptov. Vznik komunikacijega protokola poskušamo doseči skozi igranje iger imenovanja med agentoma, pri katerih agenta izmenično poimenujeta objekt na sliki in se učita besed sogovorca. Po s50.000 odigranih igrah prideta do stabilnega besedišča pri imenovanju, njuna napoved vidnega vhoda na podlagi prejetih besed pa ni povsem zanesljiva. Čeprav nepopolni, so rezultati obetavni za nadaljnje raziskave v smeri uporabe napovednega procesiranja in kažejo na zmožnost modeliranja vznika komunikacije na podlagi samoorganizacije vsakega udeleženca v pogovoru.Communication among agents is of vital importance for their effective collaboration and the development of language probably played a key role in human technological advancement. This work aspires to develop a self-organized communication system between two agents in a computer simulation in a way that is theoretically based in neuroscientific research. Language is understood in terms of a communication protocol that couples the participants of the conversation into a larger system in which the optimal operation is attained via the optimization of individuals. Agents in the simulation are designed on the principle of predictive processing, which postulates that in order to understand their environment agents must minimize the prediction error between their predictions about the future states and actual sensory input. In this way, they minimize surprisal of their sensory input and consequently improve their generative model of the world. A simple simulation is set up with two agents who are presented the same image and are equipped with the ability to form messages about it. Agents are modeled as a perceptual and generative neural network on each of the two modalities at hand, visual and verbal, with a shared internal representation of concepts. The emergence of a communication protocol is achieved by playing so-called naming games in which agents alternately name the object in the image and learn the words of their co-speaker. After 50.000 iterations of naming games, a stable vocabulary for object naming is achieved, but agents’ predictions of visual input based on the received messages is not completely reliable. While imperfect, the results show promise for future research in the direction of using the principle of predictive processing and indicate that modeling the emergence of language based on self-organization of individual participants is a viable approach