The Hallucination Effect

Is the sensitive dependence on the initial conditions in which a small change in one state in the system can result in large differences in a later state. It’s some-how connected to “The Butterfly Effect” & “The Domino Effect”.

Observing this phenomenon in Tesseract 4.x

If your training text frequently includes a certain element in a certain form.

• Example 1: A word is frequently in the Capital form Word, then when you use that trained model to recognize word, it’s going to hallucinate and see it as Word.

• Example 2: Your training text frequently includes a Space at the beginning of your sentences or at the end. Might result in slow training, non-convergence & even model corruption.

Conclusion

Mostly, The Hallucination Effect is a product of the texts that have been used to create the traineddata model.

To combat such effect, one should refrain from excessive use of a single element/form/character/ etc…

Also, one should use a vast and diverse text input for training, while cleaning and removing the parts that you deem unnecessary, because Neural Networks also learn patterns & language behavior.