What is ChatGPT?

ChatGPT describes a so-called Chatbotwhich is based on artificial intelligence can interact with people. Chatbots are basically able to establish communication between a human and a machine.

ChatGPT was published by its developer OpenAI in November 2022 and is considered the successor to the InstructGPT models. OpenAI is an American company that researches artificial intelligence and is supported by Elon Musk and Microsoft, among others. The non-profit organisation, founded in 2015, also published the GPT-2 and GPT-2 language modules, among others. GPT-3 and the programme DALL-E and its successor, DALL-E 2, which are capable of Machine learning to create images on the basis of text descriptions.

How does the language model work?

While ChatGPT is traded as a sister model of the aforementioned InstructGPT, the algorithm is built on a Model of GPT-3, specifically the GPT-3.5 series.

The language model uses what is known as "reinforcement learning from human feedback (RLHF)", whereby the foundations of the model are laid through supervised learning (supervised learning) are to be laid. For this purpose human trainers used to Training data by taking on the role of both the user and the AI assistant.

In the second step, they assisted in the creation of reward models for reinforcement learning (reinforcement learning) of the model by evaluating the responses generated by the trainers. Based on this, the Reward models through proximal policy optimisation be refined.

Online access

ChatGPT can currently be downloaded from the OpenAI website can be called up and used. After registration by means of an OpenAI account and successful login, the model can be currently be used and tested free of charge during the so-called "research preview.

OpenAI hopes to receive feedback from users at this stage, as well as user testing of the tool's strengths and weaknesses. The user agreements clarify that the language model may not be used for purposes that infringe on the rights of individuals to discover source code, develop other large-scale models that compete with OpenAI, or declare the data output to be human-generated when it is not.

The language model is designed to communicate with users in dialogue format. It should also be able to answer follow-up questions correctly within a conversation. This is possible because ChatGPT stateful is and is Reminds you of previous promptsThis allows the user to refer to it and it is understood by the language model.

ChatGPT should also be able to reject inappropriate and illegal requests and refuse replies. Limitations in the function the company states in that way by pointing out that the chatbot sometimes generates plausible-sounding but wrong and nonsensical answers. The causes of this behaviour are discussed and justified with the fact that during reinforcement learning there is no source of truth, in supervised learning the knowledge of the human trainer is decisive, and a conservative or more cautious answer policy leads to questions remaining unanswered although the system could answer them correctly. Furthermore, slight changes in the input can lead to a change in the output answer or, in the case of ambiguous queries, the model tries to guess and answer the intended question instead of asking a query.

ChatGPT often provides very extensive responses, as these have been preferred by the trainers and are therefore rewarded more highly. Although the language model is trained to prevent inappropriate requests, this cannot be entirely prevented.


What is a chatbot?

A chatbot is a dialogue system that can establish communication between a human and software. The way it works is that a human enters a message as text input or by means of spoken language and the chatbot responds to it by means of a meaningful answer. In this form, a conversation is to be established in which the bot responds to the message. Databases or through machine learning or Deep Learning is trained and can thus optimise the responses.

To strengthen the appearance of communication with a real counterpart, chatbots are often used together with so-called avatars. This is a virtual person or identity with a photo, name and fictitious human characteristics.

Basically, a distinction is made between 2 types of chatbots. Those that allow free text input and those that are based on rules. In rule-based bots, input options are suggested which can be selected in order to establish communication. Free-text bots, on the other hand, allow free input and are based on the so-called "free text". Natural Language Processing (NLP). This is a method in which machines can understand and interpret human language.

The difficulty lies in the correct interpretation of grammar and the challenge of correctly understanding the meaning and context of sentences. Machine learning is used for this case, whereas this is not necessary when using rule-based bots.

Programming and technical basics

Chatbots can be written in many different programming languages, such as PythonJava or PHP. With a library called "Chatterbot", Python offers a prefabricated framework construct in which Training data are stored in so-called corpora, which bots use for independent learning.

In addition to the possibility of setting up a chatbot on your own, there are also a variety of ready-made open source chatbot development frameworks such as "Microsoft Bot Framework", "IBM Watson" or "Botpress", which can often be created without programming knowledge and also offer interfaces to social services as well as analysis options.

All solutions (whether self-programmed or assembled using an open source modular system) are based on the same principle, responding to a human input with an output with the highest confidence level of the available data and incorporating the quality of the response for subsequent conversations.

Areas of application for chatbots

The areas of application for chatbots are very wide-ranging. For example, they can be used in the following areas:

  • Helpdesk: In this use case, the customer interacts with the chatbot to request a Answer a problem or question. In the first step, this method is intended to replace contact by e-mail or telephone and offers the customer the advantage that, in the event of success (around the clock), he or she is presented with an immediate solution or answer. If the conversation with the bot is unsuccessful, it is often possible to contact the customer via conventional means. Through automatic data collection, the bot learns with each question from a customer and can output this learning in the future.
  • Purchase advice/product recommendation: In online shops, this type of bots, also called Called a service bot, often used to offer "personal" advice based on individual preferences. The bot filters suitable products or services based on the questions asked and answers given and offers them to the customer. In a way, this is intended to correspond to personal advice on the internet and to collect data on preferences and trends through immediate customer feedback.
  • Internal application in the "back officeCompanies can also use bots within their own company without external customer contact. Here they are suitable for example as a contact point for internal questions and also for classifying and categorising messages, contributions or applications. The distribution of messages to responsible employees can also be taken over by bots.

Computational Neuroscience

What is Computational Neuroscience?

Computational neuroscience is used to research how nervous systems process their information. The computer-aided simulation of the nervous system and a lifelike representation of processes in the brain form the basis of computational neuroscience. The processing of sensory impressions is observed.

Neuro-researchers develop mathematical models based on experimentally obtained data, which are finally simulated by neuronal functions on the computer. In the process, predictions of models for neuronal behaviour are experimentally tested and optimised.

Many innovative technologies already benefit from research successes in this field. Knowledge about brain functions can now be used to design intelligent technical aids, such as driver assistance systems, self-learning computers, Robot and also intelligent prostheses. A central goal is that computer models help to recognise and heal brain malfunctions and the causes of diseases at an early stage. Therapeutic approaches can be tested virtually with the help of computational neuroscience and thus help in the continuous further development of real therapies and studies.

What models and foundations does computational neuroscience rely on?

The foundation of computational neuroscience builds on the further development of Artificial intelligence and the model of the artificial neural networks. Artificial neural networks (ANNs) can be seen as mathematical replicas of stimulus processing in the brain. These replicas are interconnected artificial neurons. Instead of electrical or chemical signals from the biological systems, algorithms with numerical values are now processed.

This system is based, for example, on machine vision on. The mathematical modelling is derived from the findings of the Neurosciencebiophysics and the theory of dynamic and complex systems. Because of their complexity, such models can usually only be simulated with the aid of computers. Experimental data often form the basis of these calculations, such as the electrophysiological properties of nerve cells and synapses and the network structures in real nerve networks.

Computational Statistics

What are Computational Statistics?

The Computational Statistics field is the point of contact between information technology and statistics. Behind the term is an essential area of Data Sciencewhich currently enjoys a great deal of attention in a wide variety of application fields and will certainly continue to do so in the future, be it for Google PageRank, spam filters in e-mail inboxes or in the context of Big data analyses.

In addition to data science, computational statistics is also subordinate to simulation science; this generally involves recreating experiments in order to minimise the amount of work involved in research or to make experiments possible in the first place.

Computational statistics is often equated with statistical computing. In fact, the former is mainly about implementing algorithms in applications; in statistical computing it is the other way around and concepts from computer science are applied to statistics.

Important methods:

  • The Markov chain is a stochastic process that is used in a wide variety of fields: Economists use it to optimise traffic systems, in financial mathematics it is used to model share prices and online marketers use it to create texts; even the popular board game Monopoly can be understood as a Markov chain. In simplified terms, this mathematical method looks at the development of random systems over time. In other words, a sequence of dice rolls whose respective dice result is, of course, independent of the previous dice roll. Using Monopoly as an example, this process could now be used to determine how likely certain game scenarios are.
  • The Monte Carlo simulation makes it possible to carry out statistical studies that would be impossible or very costly in other ways. If, for example, the average height of a person is to be determined, one could measure all the citizens of the earth and divide the sum by the world population - this is an impossible undertaking. In the Monte Carlo simulation, a smaller number of people are randomly selected, which keeps the workload low. The more measurements are made, the closer one gets to the real result - the reason for this is the law of large numbers. Monte Carlo simulation is also used in many areas: climate models predict the weather, for example, companies use it to weigh up risks and production processes in manufacturing plants are optimised with the help of this method.
  • The Maximum likelihood method is a universally applicable estimation method - in bioinformatics it is considered a standard procedure. Like Monte Carlo simulation, the maximum likelihood method is used to keep the effort as low as possible. This means: If you want to try out different parameters for a statistic, but there are no measurements for them, the maximum likelihood method is used to determine the parameter that most likely leads to the desired result.

What role do computational statistics play in the development of new technologies?

Computer-aided statistics is made up of various components. Based on the mathematical principles of probability, distribution, estimation and inference, methods (such as the Markov chain) are used to process data. Those who work in this field have mastered the procedures of statistics and their digital implementation.

In the future, work with computational statistics will play a role more than ever. Especially Areas of the Digitisation are mostly supplemented by computer-assisted statistics. In the field of autonomous driving For example, there is an urgent need for statistics; as safety is the primary concern in public road transport, computerised statistics are essential. The Nanotechnology and the medical sector in general will continue to rely on methods such as the maximum likelihood method to conduct research on DNA threads.

The Fields of technologisation require analysis by computer-aided statistics, be it the virtual reality, blockchain or the artificial intelligence.

An example of computerised statistics in the development of new technologies is an online platform for rental flats. Since the company was founded, there was the problem of the countless variables that make it difficult for landlords to set prices. Therefore, from the beginning, they relied on Data Science to calculate price suggestions for their clients. These suggestions reduce the workload for the landlord and thus make it less difficult to place an ad for the vacant flat. In turn, the resulting increase in turnover is processed statistically. Computer-aided statistics is closely interwoven with the development of new technologies; this can be seen in this example.

Condition monitoring

Condition monitoring refers to the continuous monitoring of the condition of equipment by sensors. The aim is to detect changes that indicate possible damage so that repairs can be carried out before failure occurs.

The main idea is to take action when components of a machine show certain behaviours that usually lead to a downward trend in product quality, deterioration of the machine or failure. It indicates the need for maintenance action before the actual damage occurs or the product quality changes drastically. The basis of this approach is the collection of a large amount of machine data and the methods used to detect anomalies.